Event
AGT-GSPO trains embodied VLMs to achieve error-recovery replanning
1 reports · 1 sources
Researchers including Zhuoran Wang proposed the verifiable interactive environment ERR-Env, the ERR-Bench benchmark, and the online reinforcement learning method Action-Grouped Turn-GSPO (AGT-GSPO), for training the 2B embodied vision-language model ERR-VLM to replan and recover after execution errors. The paper states that AGT-GSPO improves ERR-VLM's strict success rate for error-free completion by 10.25 percentage points and its recovery rate by 22.95 percentage points. ERR-Env provides feedback on task progress and execution errors; ERR-Bench contains held-out tasks covering normal, intermediate, and error states, and is used to evaluate recovery capability and execution reliability. AGT-GSPO combines per-turn task-progress and error-state rewards with action-grouped advantage estimation.
Report timeline
- AGT-GSPO trains embodied VLMs through verifiable interaction to achieve execution error recovery
The paper proposes the ERR-Env verifiable interaction environment, the ERR-Bench benchmark, and an online reinforcement learning method, Action-Grouped Turn-GSPO (AGT-GSPO), to train the 2B embodied VLM ERR-VLM to learn replanning and recovery after execution errors.
arXiv cs.RO ·